Prediction of microstructure evolution at the atomic scale by deep generative model in combination with recurrent neural networks

Abstract

A novel method to predict multi-atom cooperative phenomena at atomic scale is proposed based on a deep generative model in combination with recurrent neural network. The variational autoencoder (VAE) model successfully identifies three different crystal orientations in the polycrystalline nickel by using 10-dimensional latent variables and restores the image of atomic configurations reflecting the crystal orientation of each grain. Moreover, microstructure evolution of the polycrystalline iron is successfully predicted through three steps: dimensionality reduction of atomic coordinates from the MD simulation using the encoder, time evolution of latent variables using the long short-term memory (LSTM) model, and data restoration using the decoder. We successfully predict the microstructure that cannot be reproduced on the time scale of MD simulations by decoding latent variables in the future time from the LSTM model. This is a new attempt of acceleration of the MD simulation that differs significantly from conventional acceleration methods.